How we validate
Every simulated behavior is checked against a specific line in a real, open-source implementation — pinned to an exact version, never paraphrased from memory. When a number matters, it comes from the source constant or the published equation, cited. Here are the receipts.
Pinned sources
| Implementation | Pinned version | Used for |
|---|---|---|
| Linux kernellive | v6.12.93 | TCP/IP, congestion, timers |
| FRRoutinglive | frr-10.6.1 | BGP / OSPF |
| BoringSSLroadmap | eada60b75448fb37811de4c88cd1685fc90b5069 | TLS 1.3 |
| OpenSSLroadmap | openssl-3.5.7 | PKI / X.509 |
| strongSwanroadmap | 6.0.7 | IPSec / IKEv2 |
| Chromiumroadmap | not yet pinned | Browser rendering |
A worked example
Drag packet loss up in the TCP simulator and the window is cut to 70%, not the textbook 50%. That 0.70 isn't a guess — it's CUBIC's multiplicative-decrease factor, β = 717/1024 ≈ 0.7, read straight from the kernel:
# net/ipv4/tcp_cubic.c:50 @ v6.12.93
static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */
The simulator's congestion cascade cites this exact file and line, and the equilibrium window it reports uses the Mathis model W = 1.22/√p (Mathis et al., 1997) — the same equation, bounded by the bandwidth-delay product. Nothing on the screen is a plausible-looking round number.
Where every number comes from
Tier 1
A published equation or a real source constant, cited to the file:line at its pin (or the RFC/paper). CUBIC β, the RTO formula (RFC 6298), Mathis throughput.
Tier 2
Where no canonical equation exists (e.g. how many HTTP/2 streams a stall affects), a calibrated model — always labeled as a model, never presented as literal source, and registered in our internal reference ledger.